Where Should Control Reside in Multi-Agent Language-Model Systems?

Abstract

As language model agents change from simple assistants to independent systems that can collaborate and use tools, an important design question arises: where should control and oversight ( i.e. governance) be placed in these systems? Governance refers to the methods that guide how agents behave, manage information flow, and enforce operational policies. Its placement - whether centralized or distributed - directly affects the system’s safety, transparency, and runtime performance. It also impacts the ability to create formal safety arguments, which are increasingly important for using complex AI in real-world situations. While many efforts focus on aligning agents or using safety tools, there is still limited research on how different governance placements - centralized, distributed, or hybrid - affect system safety and performance throughout their lifecycle. This paper addresses that challenge by defining and comparing three governance structures for multi-agent language model systems. We examine centralized control through a single coordinator, distributed governance within individual agents, and a hybrid model that combines global oversight with local independence. These models are tested using a multi-agent platform created for open-ended question answering, which requires agents to retrieve, reason, and work together with various and unpredictable data. We looked at system behavior across several important areas: such as task completion, answer helpfulness, answer relevancy, transparency, retrieval confidence, and average runtime. The results show clear trade-offs. Distributed governance improves transparency and makes it easier to follow agent reasoning, but it also leads to longer runtimes due to additional checks and retries. Centralized and hybrid designs provide similar output quality but operate much more efficiently. To our knowledge, this is the first direct comparison of governance placement in multi-agent LLM systems. The evidence shows that governance is not a minor detail; it is a key design choice that impacts system safety, speed, and reliability in real-world tasks.